Beyond Bayesian Inference: The Correlation Integral Likelihood Framework and Gradient Flow Methods for Deterministic Sampling Authors: Piotr Gwiazda, Alexey Kazarnikov, Anna Marciniak-Czochra, Zuzanna Szymańska Published: 2025 ABSTRACT Calibrating mathematical models of biological processes is essential for achieving predictive accuracy and gaining mechanistic insight. However, this task remains challenging due to limited and noisy data, significant biological variability, and the computational complexity of the models themselves. In this method’s article, we explore a range of approaches for parameter inference in partial differential equation (PDE) models of biological systems. We introduce a unified mathematical framework, the Correlation Integral Likelihood (CIL) method, for parameter estimation in systems exhibiting heterogeneous or chaotic dynamics, encompassing both pattern formation models and individual-based models. Departing from classical Bayesian inverse problem methodologies, we motivate the development of the CIL method, demonstrate its versatility, and highlight illustrative applications within mathematical biology. Furthermore, we compare stochastic sampling strategies, such as Markov Chain Monte Carlo (MCMC), with deterministic gradient flow approaches, highlighting how these methods can be integrated within the proposed framework to enhance inference performance. Our work provides a practical and theoretically grounded toolbox for researchers seeking to calibrate complex biological models using incomplete, noisy, or heterogeneous data, thereby advancing both the predictive capability and mechanistic understanding of such systems. Introduction In recent years, biology and medicine have been undergoing a fundamental transformation from qualitative, data-poor sciences to quantitative, data-driven disciplines. Advanced technologies have enabled the study of biological systems with unprecedented precision, allowing researchers to collect vast amounts of data. In some areas, such as genome-wide transcription studies, datasets capture nearly all cellular responses to specific stimuli. Nonetheless, data availability remains limited in other instances, offering only dynamic information on protein levels or static snapshots of the spatial structure of solid tumours. Broadly speaking, two modelling paradigms are emerging, built on contrasting methodologies. Statistical and machine learning methods dominate the first paradigm, typically requiring no prior knowledge of the underlying system, as data-driven hypotheses are inferred directly from the data. In contrast, the second paradigm relies on mechanistic models, which require substantial prior knowledge of the system to manually construct representations that capture its essential dynamics. While this approach offers a significant advantage by explaining underlying mechanisms and providing a foundation for controlling the system, it also poses major challenges, as it involves many qualitatively distinct research methodologies. Mathematical models should, of course, be analytically well-posed, demonstrating properties such as existence, uniqueness, and regularity of solutions. While these properties are important, they are not sufficient for real-world applications. Appropriate illustration of the results, achieved through numerical simulations, is also necessary. These simulations not only facilitate interdisciplinary communication but also enable meaningful comparisons between numerical results and experimental observations. A critical assumption in simulations is that all parameters are known. But is this truly the case? Clearly, it is not, as for many parameters, direct experimental measurement is not possible due to physical unfeasibility or the lack of established experimental protocols. Moreover, even when data collection is feasible, it is often prone to error, yielding probability distributions rather than precise values. Consequently, some parameters must be inferred from indirect measurements, leading to the inverse problem of parameter identification. Once again, it becomes evident that practical applications give rise to theoretical challenges. Investigating the regularity of solutions with respect to the model parameters provides information on how variations in parameters lead to changes in the solution of the PDE. If small changes in parameters have a significant impact on solutions, then the inference process may become highly sensitive to noise in the observational data. Conversely, if the model exhibits robustness to variations in parameters, the inference process will be more stable and reliable (Bianchini and Colombo 2002; Dębiec et al. 2025). As a reward for that effort, we obtain parameter estimation and model validation, which indicate whether the model accurately represents the real process. A thorough understanding of the mathematical structure of the model and its properties, such as the regularity of solutions with respect to model parameters, is also necessary for another key step in developing truly predictive models, which is the construction of reliable numerical algorithms (Gwiazda et al. 2023). More precisely, the numerical scheme should satisfy the same stability estimates as the analytical solution of the equation. Finally, effective algorithms in Bayesian statistics, as well as in machine learning, require improvements in sampling methods, which involve the selection of representative data points from a larger data set to estimate the characteristics of the entire population. In the context of parameter estimation, the problem of sampling a probability distribution is of particular interest. Employing methods grounded in gradient flows in the space of probability measures and integrating them into a biological inference framework emerges as a promising direction (Ambrosio et al. 2008). The above sketch roughly illustrates the number of challenges scientists face when performing a formal model calibration. It is therefore understandable why most models of biological dynamics, especially in spatially or physiologically structured contexts, have traditionally relied on theoretical assumptions rather than empirical calibration. However, perhaps due to the increasing availability of computational power needed to perform calibration, this situation is beginning to change. Recent studies demonstrate that data-driven model calibration can reveal novel insights into biological mechanisms and is an essential prerequisite for these models to function as genuine tools for model-based data analysis. Exemplary studies span both discrete and continuous frameworks. In 2024, approximate Bayesian computation was used to calibrate single- and biphasic agent-based tumour growth models, enabling the identification of key parameters such as proliferation rates and spatial constraints on cell division that distinguish tumour growth phases (Wang et al. 2024). In a paper published the same year, the identifiability of several Fisher–KPP variants for cell invasion was assessed using profile likelihoods, revealing that more complex models, while biologically richer, are often less identifiable and require more data for reliable parameter estimation (Liu et al. 2024). A year earlier, Bayesian calibration was applied to parameterise a multiphase porous-media tumour model, combining global sensitivity analysis with Bayesian methods to indirectly characterise mechanical properties of neuroblastoma spheroids that are otherwise difficult to measure (Hervas-Raluy et al. 2023). In 2023, approximate Bayesian computation and gradient matching were compared in a PDE model of cancer spread, demonstrating that approximate Bayesian computation (ABC) is more robust to measurement noise and therefore better suited for parameter recovery in noisy biological data (Xiao et al. 2021). In the same year, calibrated diffusion–proliferation coefficients were investigated in a murine glioma model, helping to explain the large inter-subject variability in tumour growth by testing hypotheses about diffusion and proliferation dynamics (Lima et al. 2021). Earlier, in 2017, calibration and validation of Gompertzian spheroid growth were performed, demonstrating a practical workflow for uncertainty quantification and improving predictive accuracy under noisy data conditions (Collis et al. 2017). Notably, even elementary mechanistic models, when systematically calibrated to data, can yield novel biological insights. For instance, in adult neurogenesis, mathematical models calibrated to experimental data on the dynamics of distinct cell subpopulations have challenged prevailing hypotheses by showing that the age-related decline in neurogenesis is primarily due to reduced activation rates of quiescent neural stem cells rather than changes in their self-renewal probabilities, thereby revising the understanding of how stem cell dynamics change with ageing (Kalamakis et al. 2019; Ziebell et al. 2018). Together, these studies underscore the transformative potential of parameter identification, not only for improving predictive accuracy but also for uncovering hidden mechanisms and reshaping biological understanding. This paper discusses the potential of new mathematical methods to address a key challenge in the practical application of mathematical models: the lack of proper calibration. It outlines the core ideas behind various sampling techniques, ranging from well-established methods such as the Metropolis and Metropolis-Hastings algorithms to more advanced approaches designed to mitigate the limitations of existing solutions. Building on this foundation, we present the recently developed Correlation Integral Likelihood (CIL) method as a unifying framework tailored for systems characterised by spatial heterogeneity, stochasticity, or chaotic behaviour. We also discuss how deterministic gradient flow methods, when integrated within the proposed framework, may enhance the efficiency and accuracy of inference. In summary, we present a practical and theoretically robust toolkit for calibrating complex biological models using incomplete, noisy, or heterogeneous data by combining stochastic and deterministic inference strategies within the CIL framework, thereby providing a flexible approach to parameter estimation that enhances both predictive power and mechanistic understanding. To guide the reader, the remainder of this article is organised as follows. In Section 2, we review the foundations of parameter identification formulated as inverse problems, highlighting the challenges that arise in complex biological systems. Section 3 introduces our CIL framework, beginning with its application to chaotic dynamical systems (Subsection 3.1) and extending to spatially extended systems, including pattern formation and agent-based models (Subsection 3.2). Section 4 explores sampling strategies for parameter inference, comparing stochastic Markov Chain Monte Carlo (MCMC) methods (Subsection 4.1) with deterministic gradient flow approaches (Subsection 4.2). Finally, Section 5 discusses the implications of our findings and outlines future directions for advancing inference methods in mathematical biology. Parameter Identification via Inverse Problems Inverse problems can be tackled either with deterministic methods, which minimise a predefined cost function, or within a broader stochastic framework, such as Bayesian methods, in which model parameters are treated as random variables described by probability distributions. The key concept is the posterior distribution, which represents the conditional distribution of the parameters given the observations, allowing inferences to be made about individual parameter values. By Bayes’ theorem, it is given by1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \pi (\theta |D)=\frac{\pi (\theta )\ell (D|\theta )}{\int _{\Theta }\pi (\theta )\ell (D|\theta )d\theta }\,, \end{aligned}$$\end{document}π(θ|D)=π(θ)ℓ(D|θ)∫Θπ(θ)ℓ(D|θ)dθ,where D denotes the collected data, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}θ represents a given vector of parameters, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Theta $$\end{document}Θ is the space of all parameters. The posterior probability density, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi (\theta \mid D)$$\end{document}π(θ∣D), represents the probability density of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}θ given the data D. The prior probability density, denoted by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi (\theta )$$\end{document}π(θ), is specified based on some prior knowledge, while the likelihood function, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ell (D \mid \theta )$$\end{document}ℓ(D∣θ), quantifies the probability of observing the data D given the parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}θ. Suppose we are given a theoretical model that describes the dynamics of a process using a system of differential equations. The objective is to estimate an unknown parameter vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta \in \mathbb {R}^d$$\end{document}θ∈Rd for this model from observed data \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$D \in \mathbb {R}^k$$\end{document}D∈Rk, within a probabilistic framework grounded in Bayes’ theorem. More precisely, under certain assumptions, the relationship between parameter vector \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}θ and data D is given by:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} D = \mathcal {M}(\theta ) + \eta , \end{aligned}$$\end{document}D=M(θ)+η,where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}: \mathbb {R}^d \rightarrow \mathbb {R}^k$$\end{document}M:Rd→Rk is the forward operator that maps the parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}θ to the observation space, while \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta $$\end{document}η represents the observational noise. We assume that the prior probability density \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi (\theta )$$\end{document}π(θ) follows a Gaussian distribution, that is, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta \sim \mathcal {N}(m, \Sigma )$$\end{document}θ∼N(m,Σ). Under this assumption, the prior density is given by3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \pi (\theta ) = e^{-\frac{1}{2} |\theta -m|^2_{\Sigma }}. \end{aligned}$$\end{document}π(θ)=e-12|θ-m|Σ2.The term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\frac{1}{2} |\theta - m|^2_{\Sigma }$$\end{document}12|θ-m|Σ2 represents prior knowledge about \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}θ, penalizing deviations from the prior mean m, scaled by the prior covariance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Sigma $$\end{document}Σ. Now, suppose that the noise \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\eta \sim \mathcal {N}(0, \Gamma )$$\end{document}η∼N(0,Γ) represents measurement uncertainty. Then the likelihood of observing D given parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}θ is4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \ell (D|\theta ) = e^{-\frac{1}{2} |D - \mathcal {M}(\theta )|^{2}_{\Gamma }}. \end{aligned}$$\end{document}ℓ(D|θ)=e-12|D-M(θ)|Γ2.Similarly, it penalises deviations between the model predictions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {M}(\theta )$$\end{document}M(θ) and the actual observations D, scaled by the noise covariance \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\Gamma $$\end{document}Γ, where5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} |D - \mathcal {M}(\theta )|^{2}_{\Gamma } = (D - \mathcal {M}(\theta ))^T \Gamma ^{-1}(D - \mathcal {M}(\theta )). \end{aligned}$$\end{document}|D-M(θ)|Γ2=(D-M(θ))TΓ-1(D-M(θ)).The above leads to the posterior density \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi (\theta |D)$$\end{document}π(θ|D), which takes the form6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \pi (\theta |D) \propto e^{-V(\theta )}, \quad \text {where} \quad V(\theta ) = \frac{1}{2} |D - \mathcal {M}(\theta )|^{2}_{\Gamma } + \frac{1}{2} |\theta - m|^2_{\Sigma }. \end{aligned}$$\end{document}π(θ|D)∝e-V(θ),whereV(θ)=12|D-M(θ)|Γ2+12|θ-m|Σ2.To demonstrate the procedure of parameter identification using the inverse problem approach, we apply it step by step to a specific example: the mathematical model of cancer colony dynamics studied in Szymańska et al. (2021); Gwiazda et al. (2023), which is based on in vitro experiments (Folkman and Hochberg 1973). In these experiments, the growth of multicellular spheroid cultures was monitored by measuring their diameters at discrete time points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_i$$\end{document}ti. The proposed mathematical model, formulated as an integro-differential equation, captures the dynamics of cancer cell colony expansion. The objective is to calibrate the model with experimental data and develop a quantitative tool capable of reliably predicting colony growth. Consider7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \partial _t n(x,t) \ =\ \alpha \, k* n(x,t) \,\Bigl ( 1 \ -\ n(x,t) \Bigr ), \end{aligned}$$\end{document}∂tn(x,t)=αk∗n(x,t)(1-n(x,t)),where n(x, t) denotes the density of cancer cells at time t and location x. The parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\alpha $$\end{document}α represents the proliferation rate, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k = k(x)$$\end{document}k=k(x) is a time-independent, radially symmetric kernel function with compact support, ensuring that\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} k *n(x,t) \ =\ \int _{\mathbb {R}^3} k(x-y) n(y,t)\;dy. \;\; \end{aligned}$$\end{document}k∗n(x,t)=∫R3k(x-y)n(y,t)dy.It is assumed that k(x) has a profile K, meaning that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$k(x) = K(|x|)$$\end{document}k(x)=K(|x|) for \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x \in \mathbb {R}^3$$\end{document}x∈R3. The normalised characteristic function of a ball is adopted as the kernel profile8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} K\big (|x|\big ) = \frac{3}{4\pi } \sigma _k^{-3} \, \mathbbm {1}_{\left[ 0,\sigma _k\right] }\big (|x|\big ), \end{aligned}$$\end{document}K(|x|)=34πσk-310,σk(|x|),where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _k$$\end{document}σk represents the kernel size (radius). To highlight the core idea, the initial condition is simplified relative to the original article. Specifically, it is assumed to be the characteristic function of a ball, where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _s$$\end{document}σs denotes the colony radius at the start of the measurements. Therefore, we take9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} p(r,0) = 4 \pi r^2 \mathbbm {1}_{\left[ 0,\sigma _s\right] }(r). \end{aligned}$$\end{document}p(r,0)=4πr210,σs(r).The inverse problem approach requires linking the model equations to the observed data-specifically, in our case, diameter (or equivalently, radius) measurements, and typically involves extensive simulations. We exploit the characteristics of the experimental data and the assumed radial symmetry of the kernel to reformulate the model in spherical coordinates (Szymańska et al. 2021; Gwiazda et al. 2023). This transformation reduces the problem’s dimensionality, yielding a theoretical framework in which \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$p_{\theta }(r, t)$$\end{document}pθ(r,t) describes the radial density dynamics of the colony, with r(t) representing the colony radius, t denoting time, and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}θ being a vector of model parameters, that is, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta = [\alpha , \sigma _s, \sigma _k]$$\end{document}θ=[α,σs,σk]. An additional benefit of this change of variables is improved computational efficiency. While the concept of a radius is intuitive, it must be precisely defined for parameter estimation, especially in the context of Eq. (7), which does not preserve compact support. Measurements of spheroid diameters, and thus their radii, inevitably contain errors. Assuming a homogeneous mass distribution and that the colony radius defines a sphere enclosing 95% of the total mass, we derive the following expression10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {M}(\theta ) = r_{\theta }(t) = \inf \left\{ q: \int _{0}^q p_{\theta }(r,t)\,dr > 0.95 \cdot \int _{0}^{\infty } p_{\theta }(r,t)\, dr\right\} , \end{aligned}$$\end{document}M(θ)=rθ(t)=infq:∫0qpθ(r,t)dr>0.95·∫0∞pθ(r,t)dr,where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\int _{0}^{\infty } p_{\theta }(r,t)\,dr$$\end{document}∫0∞pθ(r,t)dr corresponds to the colony mass at time t. Another modelling choice is the error model. We assume deviations in measured spheroid diameters scale proportionally to their true values, which is a multiplicative error model, given by11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} r_{o}^i = r_{\theta }(t_i) \cdot Z_i, \;\;\; \text {with} \;\;\; \log (Z_i) \sim N(0,\sigma _{o}^2), \end{aligned}$$\end{document}roi=rθ(ti)·Zi,withlog(Zi)∼N(0,σo2),or alternatively12\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \log (r_{o}^i) = \log (r_{\theta }(t_i)) + \tilde{Z}_i, \;\;\; \text {with} \;\;\; \tilde{Z}_i \sim N(0,\sigma _{o}^2), \end{aligned}$$\end{document}log(roi)=log(rθ(ti))+Z~i,withZ~i∼N(0,σo2),where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_o^i$$\end{document}roi is the measured colony radius at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_i$$\end{document}ti, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_{\theta }(t_i)$$\end{document}rθ(ti) is the actual colony radius at that time given by Eq. (10), and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _o$$\end{document}σo represents the measurement error, assumed to be homogeneous over time. As the next step in the procedure, we extend the parameter vector associated with the theoretical model Eq. (7) by including the parameter \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\sigma _o$$\end{document}σo, yielding \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta := \theta \oplus \sigma _o$$\end{document}θ:=θ⊕σo. The resulting vector of unknown parameters is \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta = \big [\alpha , \sigma _{k}^2, \sigma _{s}^2, \sigma _{o}^2\big ]$$\end{document}θ=[α,σk2,σs2,σo2], where each component corresponds to the proliferation rate, kernel size, initial colony radius, and measurement error, respectively. A direct consequence of assumption Eq. (11) is that the likelihood function \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\ell (D \mid \theta )$$\end{document}ℓ(D∣θ) is given by13\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \ell (D \mid \theta ) = \prod _i \frac{1}{\sqrt{2 \pi }\sigma _{o}} \exp {\left( \frac{-\Big (\log (r_{o}^i)-\log (r_{\theta }(t_i))\Big )^2}{2 \sigma _{o}^2}\right) }, \end{aligned}$$\end{document}ℓ(D∣θ)=∏i12πσoexp-(log(roi)-log(rθ(ti)))22σo2,where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$r_{\theta }(t_i)$$\end{document}rθ(ti) is the colony radius at time \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_i$$\end{document}ti for vector of parameters \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta $$\end{document}θ. For convenience, and to avoid imposing positivity constraints, we work with the logarithms of the parameters rather than the parameters themselves. By evaluating the likelihood Eq. (13) and assuming a prior distribution \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\pi (\theta )$$\end{document}π(θ), we obtain the posterior probability density Eq. (1). However, computing the posterior can be computationally expensive, and considerable effort has been devoted to optimising the cost of its evaluation. Parameter inference in models with emergent heterogeneity A central challenge in data-driven model identification arises when limited data must be used to infer parameters in systems that exhibit solution heterogeneity. This is particularly evident in models such as pattern formation models and individual-based models, where complex spatial structures emerge from underlying dynamics. In many cases, available data are restricted to single spatial snapshots, providing limited information about the temporal processes that drive structure formation. Moreover, these heterogeneous outcomes are often highly sensitive to initial conditions, which are typically unknown or poorly constrained. For example, in two-dimensional Turing models, even small perturbations of the initial conditions can lead to markedly different patterns. Mixed-mode patterns, such as combinations of spots and stripes, may emerge from the same parameter set, reflecting the system’s sensitivity to initial states. The limited availability of data makes it especially challenging to identify and interpret such complex behaviours. Another illustrative case in mathematical biology is stochastic modelling, particularly in individual-based models, where outcomes inherently vary between simulation runs. Here, random processes govern the behaviour of individual agents or entities, so even with identical initial conditions and parameters, repeated simulations can yield different results. To address this challenge, we present a recent approach known as the Correlation Integral Likelihood method. The name reflects its foundation in the concept of correlation dimension, originally developed for characterising chaotic attractors (Grassberger and Procaccia 1983; Grassberger 2007). Rooted in nonlinear dynamics and chaos theory, the correlation dimension provides a means of estimating the fractal dimension of a point cloud generated by simulations of chaotic systems. Building on and generalising this concept (Haario et al. 2015), the CIL method constructs a stochastic cost function for parameter estimation in chaotic dynamical systems. This framework is broadly applicable to models with intrinsically random or heterogeneous outcomes. Initially developed for chaotic dynamical systems (Haario et al. 2015; Springer et al. 2019, 2021; Shah et al. 2023; Fumagalli et al. 2025), the CIL method has since been successfully extended to a variety of other model classes. These include cellular automata models (Kazarnikov et al. 2025) as well as partial differential equation (PDE) models of pattern formation (Kazarnikov and Haario 2020; Kazarnikov et al. 2023; Pan et al. 2025) and of rumour propagation, in both continuous and network-based spatial domains (Zhu and He 2022; Li and Zhu 2024; Tao and Zhu 2025). More recently, the method has been adapted to address scenarios with limited data through the introduction of a synthetic variant, known as the Synthetic Correlation Integral Likelihood (SCIL) (Kazarnikov et al. 2023). Additionally, the method has been extended to handle mixed mode patterns by constructing an ensemble of synthetic likelihoods, thereby capturing the variability of coexisting pattern types (Kazarnikov et al. 2025). The CIL approach for chaotic dynamical systems Chaotic dynamical systems are characterised by extreme sensitivity to initial conditions, where even small changes in the initial state or solver parameters can lead to drastically different trajectories, a phenomenon commonly referred to as the butterfly effect. Despite these divergent paths, the trajectories remain confined to the same underlying attractor, typically exhibiting fractal geometry. As a result, there is no single model trajectory corresponding to fixed parameters; instead, there exists an ensemble of trajectories that share the same global geometric properties. This can be understood as the system’s natural or intrinsic variability. While in the parameter estimation problem, the goal is to fit a model to observed data sampled at discrete time points, the inherent sensitivity of chaotic systems makes conventional methods, such as least squares fitting, ineffective, as they rely on pointwise comparisons between data and model predictions. Due to the exponential divergence of trajectories over time, even simulations with identical parameters will quickly diverge from one another because of the system’s intrinsic variability. As a result, an alternative metric for model comparison is required. The CIL method addresses these challenges by robustly comparing the global structure of chaotic attractors. It introduces a novel metric based on the concept of generalised correlation integrals (Grassberger and Procaccia 1983). Consider a chaotic dynamical system given by14\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \dot{x} = F(x,\theta _0), \end{aligned}$$\end{document}x˙=F(x,θ0),where \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$x \in \mathbb {R}^n$$\end{document}x∈Rn is the state vector and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _0 \in \mathbb {R}^p$$\end{document}θ0∈Rp is the vector of parameters. Suppose measurements yield two trajectories of system Eq. (14)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} S(\theta _0) = \{s_1, s_2, \dots , s_N\} \subset \mathbb {R}^n \end{aligned}$$\end{document}S(θ0)={s1,s2,⋯,sN}⊂Rnand\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \tilde{S}(\theta _0) = \{\tilde{s}_1, \tilde{s}_2, \dots , \tilde{s}_N\} \subset \mathbb {R}^n, \end{aligned}$$\end{document}S~(θ0)={s~1,s~2,⋯,s~N}⊂Rn,sampled at time points \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$t_1< t_2< \cdots < t_N$$\end{document}t1 0, \end{aligned}$$\end{document}Iψ(ϱ(t)‖ϱ∞):=-∫Rn|∇w|2ϱ∞dθ,w=Fψϱ(t)ϱ∞,Fψ(μ)=∫1μψ′′(s)ds,μ>0,The strict convexity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\psi $$\end{document}ψ implies a strict monotonicity of \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$F_\psi $$\end{document}Fψ and hence, for any probability densities \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varrho (t)$$\end{document}ϱ(t) and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varrho _\infty $$\end{document}ϱ∞, the functional \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {I}_\psi (\varrho (t) \Vert \varrho _\infty )$$\end{document}Iψ(ϱ(t)‖ϱ∞) is non-positive and equals zero if and only if \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varrho (t) = \varrho _\infty $$\end{document}ϱ(t)=ϱ∞. This expresses the fundamental property that any relative entropy \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\mathcal {E}_\psi (\varrho \Vert \varrho _\infty )$$\end{document}Eψ(ϱ‖ϱ∞) decreases in time, provided that \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\psi $$\end{document}ψ is strictly convex. Under suitable assumptions, one can also derive convergence rates using logarithmic Sobolev inequalities (Arnold et al. 2001). In practice, numerical implementation of gradient flow-based samplers often relies on simulating the Langevin equation,24\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} {\!\text {d}} X(t) = -\nabla V(X(t))\!\text {d}t + \sqrt{2}\, \!\text {d}B(t)\,, \end{aligned}$$\end{document}dX(t)=-∇V(X(t))dt+2dB(t),where B is the Brownian motion, see the seminal paper (Roberts and Tweedie 1996), as well as more recent studies (Chewi et al. 2024; Dalalyan 2017, 2016; Durmus et al. 2019; Dalalyan et al. 2022; Durmus and Moulines 2017, 2019; Erdogdu and Hosseinzadeh 2021; Erdogdu et al. 2022; Mousavi-Hosseini et al. 2023; Vempala and Wibisono 2019). Convergence analysis of Langevin-based sampling methods, rooted in the interpretation of the Fokker–Planck equation as a gradient flow in the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_2$$\end{document}W2-Wasserstein distance, has been recently presented (Durmus et al. 2019). A notable limitation of this approach is that, due to stochastic fluctuations, a large number of trajectories is required to obtain accurate approximations. To overcome this, recent efforts have focused on deterministic particle-based approximations. A major challenge in this context is that the variational structure of the Fokker–Planck equation does not extend to discrete (atomic) measures. Specifically, the density \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varrho (\textbf{r})$$\end{document}ϱ(r) satisfies a variational principle with the corresponding free energy\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \mathcal {F}[\varrho ] = \int V(\theta )\,\varrho (d\theta ) + \int \varrho (\theta ) \log \varrho (\theta ) \, d\theta . \end{aligned}$$\end{document}F[ϱ]=∫V(θ)ϱ(dθ)+∫ϱ(θ)logϱ(θ)dθ.and the entropy term \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\int \varrho (\theta ) \log \varrho (\theta ) \, d\theta $$\end{document}∫ϱ(θ)logϱ(θ)dθ becomes infinite for empirical measures composed of Dirac deltas. One popular approach to bypass this issue is Stein Variational Gradient Descent (SVGD). Instead of directly simulating the Fokker–Planck equation, SVGD evolves the empirical distribution\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \varrho ^N(t) = \frac{1}{N} \sum _{i=1}^N \delta _{\theta _i(t)}, \end{aligned}$$\end{document}ϱN(t)=1N∑i=1Nδθi(t),according to a nonlocal PDE25\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \partial _t \varrho = \nabla \cdot \big ( \varrho K * (\nabla \varrho + \varrho \nabla V) \big ), \end{aligned}$$\end{document}∂tϱ=∇·(ϱK∗(∇ϱ+ϱ∇V)),where K is a positive-definite kernel and \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$*$$\end{document}∗ denotes the convolution operator. The particle dynamics follow the system of ordinary differential equations26\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\begin{aligned} \frac{d}{dt} \theta _i(t) = -\frac{1}{N} \sum _{j=1}^N \nabla K(\theta _i(t) - \theta _j(t)) - \frac{1}{N} \sum _{j=1}^N K(\theta _i(t) - \theta _j(t)) \nabla V(\theta _j(t)), \end{aligned}$$\end{document}ddtθi(t)=-1N∑j=1N∇K(θi(t)-θj(t))-1N∑j=1NK(θi(t)-θj(t))∇V(θj(t)),with initial conditions \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\theta _i(0)$$\end{document}θi(0) given by \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varrho ^N(0) = \frac{1}{N} \sum _{i=1}^N \delta _{\theta _i(0)}$$\end{document}ϱN(0)=1N∑i=1Nδθi(0) that approximates the initial density \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\varrho (0)$$\end{document}ϱ(0). These deterministic dynamics allow for efficient sampling while preserving key properties of the underlying distribution. The convergence and stability of the particle system Eq. (26) have been recently analysed within the framework of the \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$W_2$$\end{document}W2 Wasserstein distance (Carrillo and Skrzeczkowski 2025). The analysis was also conducted with respect to the dual bounded Lipschitz distance (also known as the flat metric), a metric used for studying nonlocal growth and transport problems in the space of nonnegative Radon measures (Düll et al. 2021). The method has also been analysed in the continuous setting, where kernels were constructed that ensure exponential convergence of the solution toward the target distribution (Carrillo et al. 2024). Another sampling approach that connects Langevin diffusion with deterministic dynamics is the so-called blob method (Carrillo et al. 2019). In this context, a nonlocal equation is introduced to approximate the Fokker–Planck equation, Eq. (21). The theoretical understanding of the method remains incomplete, despite independent studies that have explored its properties (Craig et al. 2025; Carrillo et al. 2024). However, recent works have provided the first convergence rates for the blob method in a simplified case of the porous medium equation (Carrillo et al. 2025; Amassad and Zhou 2025). In summary, gradient flow-based methods offer a powerful and theoretically well-founded alternative to stochastic sampling. Their connection to optimal transport, entropy minimisation, and PDEs provides a rich framework for both analysis and implementation. As computational methods improve, such deterministic algorithms are becoming increasingly competitive for high-dimensional sampling tasks. Discussion and Outlook Calibrating complex biological models remains a central challenge in mathematical biology, where data are often noisy, incomplete, and heterogeneous, and where model dynamics may be nonlinear, multiscale, or even chaotic. Nevertheless, successful parameter inference has enabled significant progress across various biological domains. Traditional Bayesian approaches to inverse problems provide a powerful statistical framework for parameter estimation by treating model parameters as random variables and seeking to characterise their posterior distribution. Derived using Bayes’ theorem, the posterior combines prior knowledge with the likelihood of observed data, thereby allowing uncertainty to be explicitly modelled. This probabilistic treatment offers valuable insights into the inherent correlations between parameters and the reliability of the resulting estimates. Among Bayesian techniques, MCMC methods, particularly the Metropolis and Metropolis–Hastings algorithms, are widely used to approximate posterior distributions. These methods construct a Markov chain whose stationary distribution matches the target posterior by generating a sequence of samples. Starting from an initial state, typically informed by prior knowledge, the algorithm iteratively proposes new candidate states using a probability density function, often modelled as a multivariate normal distribution. The Metropolis algorithm assumes symmetry in the proposal distribution, whereas the Metropolis–Hastings algorithm allows for asymmetry, making it more general and adaptable. Despite their theoretical appeal, MCMC based techniques can be computationally demanding, especially for high-dimensional systems or models that exhibit sensitive dependence on initial conditions. In such settings, convergence can be slow, and performance becomes highly sensitive to the choice of prior and proposal distribution. As a result, standard Monte Carlo integration may become inefficient or even intractable. These challenges have led to the development of more advanced sampling strategies aimed at improving convergence rates and reducing computational cost. In some situations, experimental data contain such a limited amount of information that it becomes impossible to derive an analytical expression for the discrepancy between the model output and the experimental data, thereby preventing the use of standard Bayesian methods. These challenges can be addressed with the CIL method, which is a robust statistical inference framework, that compares full spatio-temporal solutions by encoding pattern similarity in the distribution of scalar features (pairwise distances). This enables its effective use even when data are incomplete, scaled, or non-stationary, making it particularly well-suited for biological models that generate complex heterogeneous structures. The accuracy of parameter identification can be significantly enhanced by incorporating multiple norms (Kazarnikov et al. 2023). Model-specific features can also be integrated into the approach. For example, in the case of cellular automata models, various meaningful features were used alongside distance metrics to improve estimation accuracy (Kazarnikov et al. 2025). The applicability of the CIL method extends well beyond classical Turing patterns. Its effectiveness has been demonstrated in diverse contexts, including mechano-chemical models that couple chemical processes with tissue mechanics, discontinuous far-from-equilibrium patterns produced by reaction-diffusion-ODE systems, and stochastic cellular automata. Moreover, the methodology is not restricted to stationary patterns. It remains valid for spatio-temporal structures such as chemical waves and Cahn–Hilliard phase separation dynamics. From a methodological standpoint, time-dependent dynamics of pattern formation models can be treated analogously to chaotic systems. This versatility makes the CIL method broadly applicable across a wide spectrum of spatio-temporal models. The CIL framework has so far been validated primarily on simulated datasets and, more recently, on an experimental chemical case study (the CIMA reaction), where the underlying reaction network is well characterised and the reduced model used for inference is effectively a known approximation of the true chemistry (Kazarnikov et al. 2025). Applying CIL to real biological systems constitutes the next major challenge. Biological datasets are not only typically sparse and noisy, but also the underlying processes are only partially understood. In such contexts, mechanistic models are often hypothetical, and the primary aim is not to recover parameters of a fully established model but rather to discriminate between competing modelling hypotheses in order to identify plausible biological mechanisms. A promising direction for future CIL applications is parameter estimation in individual-based models, where model outputs can be directly confronted with image-derived spatial statistics. This is particularly relevant for biological systems in which cell-level behaviours, such as division, differentiation, and migration, drive emergent tissue-scale patterns. A concrete example is provided by lattice-based models of stem cell populations, which can be compared with high-resolution imaging datasets, such as quantitative live imaging of stem cell dynamics in zebrafish neural tissues (Dray et al. 2021). Similarly, vertex- and lattice-based models of epithelial morphogenesis, generating spatial mosaics of distinct cell fates, can be confronted with high-resolution microscopy data quantifying the spatial arrangement of labelled cell populations in developing tissues (Fletcher et al. 2014; Alt et al. 2017). A further example is provided by pattern-formation models whose simulated output can be matched to in situ hybridisation images showing spatial distributions of mRNA corresponding to one of the model variables (Mercker et al. 2025). Complementing the likelihood formulation, we discussed deterministic alternatives to classical MCMC, grounded in the theory of gradient flows on the space of probability measures. By interpreting sampling as a transport problem minimising the Kullback–Leibler divergence, one can design algorithms, such as Fokker–Planck solvers or particle-based variational methods, that converge deterministically toward the target posterior. These approaches significantly reduce computational cost and variance, offering practical advantages in high-dimensional or computationally stiff inference tasks. Together, these methodological advances constitute a scalable and versatile toolbox for parameter estimation, as demonstrated on a range of models describing biological processes. On example of a mathematical model of cancer colony dynamics we illustrated the utility of Bayesian inference in linking theoretical predictions with experimental observations, while rigorously accounting for noise and measurement uncertainty. The applicability of the CIL method is showcased through models that produce heterogeneous outputs, such as Turing pattern formation systems, stochastic cellular automata, and Cahn–Hilliard-type models. A key strength of the CIL approach lies in its ability to extract meaningful features even from noisy, incomplete, or scaled data. Furthermore, modern numerical techniques such as adaptive meshing, GPU acceleration, and parallel computing enable the development of computationally efficient implementations of the method, making it particularly suitable for contemporary computational biology. Looking forward, the integration of CIL with machine-learning-based surrogates or reduced-order models has the potential to further enhance its applicability to large-scale biological systems. In turn, the gradient flow approach could provide valuable support in designing structure-preserving inference algorithms tailored to specific model classes, such as those obeying conservation laws or exhibiting symmetry-breaking bifurcations.